Inspur has launched the world’s densest and most powerful AI server AGX-2 (model name: Inspur NF5288M5) with 300GB/s high speed NVIDIA® NVLink™ to connect 8 high performance GPU accelerators within 2U space. This server is immensely important for AI training and HPC applications where it can provide 200 times higher performance than the traditional dual-CPU servers. The flexible and scalable design can achieve dual-system horizontal expansion to 16 GPU structure which extremely useful to cater for different computing scenarios.
The AGX-2 approached with amazing features for achieving the best computing performances in the AI and HPC domain. It‘s outstanding performance, incredible computing density and flexible configuration options make itself most demanding in the industry.
Extreme computing density
AGX-2 utilizes P100’s Linpack’s floating point computing power to achieve 29.33TFLOPS, 2.47 times that of NF5288M4, which also utilizes P100. When it comes to AI deep-learning model training, AGX-2, which utilizes TensorFlow framework and GoogLeNet model, processes data at 1165 images per second, can provide single node with a peak value computing power of 960 Tensor TFLOPs. Moreover, AGX-2 is based on high-density design can easily allow a 42U rack’s cluster’s peak performance to rise above 1 PFLOPS (one quadrillion floating point operations per second). It can make inter-GPU bandwidth as high as 300 GB/s and allow for little to no lag, allowing an over 60% increase in the efficiency of parallel GPU. Apart from this, it has incorporated two latest Intel® Xeon® Scalable Processor and 16 2666 MHz speeds based memory sticks.
Ultimate flexible design
AGX-2 connects CPU and GPU resources with PCIe cable, enabling flexible adjustment of the CPU connection bandwidth and the number of connections. In response to different AI applications, it’s better for PCIe resources to be allocated on demand. Flexible computing architecture allows one or two CPU to manage 8 GPU or achieve scale-up up to 16GPU by way of expanding box by GPU. PCIe I / O, 8 U.2 slots, or up to 4 network interface cards of 100Gbps InfiniBand provided by the server can flexibly adjust topology according to the calculation. The resilient heterogeneous platform of AGX-2 is enough to support a variety of AI scenes. Furthermore, it can provide point-to-point communication within the system and decrease the amount of heterogeneous communication, independent of CPU.
The AGX-2 provides utmost intelligent management strategy to achieve the industry leading performance. Provide specific Ethernet port for management and supports remote monitoring, SMTP KVM, SNMP management, Virtual Media and redundant management system.
NF5288M5 Technical Specification
Intel® C624 chipset
2 Intel Xeon Scalable processors, TDP up to 165W
Processor core available
Maximum 28-core per processor
3.6 GHz, maximum depending on processor
38.5 MB L3 cache
8 SXM2 GPU with NVLink
8 x SXM2 GPU on NVLink cube mesh with 2 x 96lane PCIe Switch，4 x PCIex16 HHHL Rear slot
8 PCIe GPU with PCIe Switch
8 x PCIe3.0x16 double width GPGPU
16 DIMM slots
RDIMMs, LDIMMs and Apache Pass
Memory protection features
Support front 8 × 2.5 inch SAS/SATA hard disk or U.2 NVMe SSD drive
Support 2 × SATA or PCIe M.2 SSD on board
SAS 3108 Mezz Card support RAID 0, 1, 5, 10
I/O Expansion slot
2* HHHL x16 card
1* SAS Mezz x8 card
Optional 4* PCIE x16 card
Integrated LAN controller; up to 4*10GbE
Integrated I/O port
2 front set USB 3.0 port, 1 VGA 1 ID button & ID LED & BMC reset button, 1 system power button & LED, 1 system reset button
2 rear set USB 3.0 port, 1 VGA, 1 serial port, 1 ID button & LED, 1 BMC management port